• Mitch Paro, William A Lambert, Nathan K Leclair, Robert Romano, Petronella Stoltz, Jonathan E Martin, David S Hersh, and Markus J Bookland.
• School of Medicine, University of Connecticut Health Center, Farmington, Connecticut, USA.
• Neurosurgery. 2022 May 1; 90 (5): 613-618.

BackgroundThe authors have developed pretrained machine learning (ML) models to evaluate neonatal head shape deformities using top-down and facial orthogonal photographs of the patient's head. In previous preliminary analysis, this approach was tested with images from an open-source data bank.ObjectiveTo determine the accuracy of pretrained ML models in identifying craniosynostosis among patients seen in our outpatient neurosurgery clinic.MethodsWe retrospectively reviewed top-down and facial orthogonal images of each patient's head and provider clinical diagnosis from the same encounters. Head shape classifications generated from 3 pretrained ML models (random forest, classification and regression tree, and linear discriminant analysis) were applied to each patient's photograph data set after craniometric extraction using a predefined image processing algorithm. Diagnoses were codified into a binary scheme of craniosynostosis vs noncraniosynostosis. Sensitivity, specificity, and Matthew correlation coefficient were calculated for software vs provider classifications.ResultsA total of 174 patients seen for abnormal head shape between May 2020 and February 2021 were included in the analysis. One hundred seven patients (61%) were seen in-person and 67 (39%) through telemedicine. Twenty-three patients (13%) were diagnosed with craniosynostosis. The best-performing model identified craniosynostosis with an accuracy of 94.8% (95% CI 90.4-97.6), sensitivity of 87.0% (95% CI 66.4-97.2), specificity of 96.0% (95% CI 91.6-98.5), and Matthew correlation coefficient of 0.788 (95% CI 0.725-0.839).ConclusionMachine learning-driven image analysis represents a promising strategy for the identification of craniosynostosis in a real-world practice setting. This approach has potential to reduce the need for imaging and facilitate referral by primary care providers.Copyright © Congress of Neurological Surgeons 2022. All rights reserved.

18 keywords...

hide…